Yolo driven computer vision framework for paddy disease

Paddy cultivation faces various disease challenges, due to these issues, around 125 tons of annual losses of rice production happen globally. This research work proposes Domain specific optimizations of YOLOv8 for paddy disease detection with integration of Paddy-Adaptive Loss Function (PALF). This design focuses on the improvement of combining focal loss, IoU-based localisation, texture consistency regularisation, and severity-weighted penalties with optimised weighting parameters, combined with a Hierarchical Feature Fusion Network (HFFN) employing attention-guided multiscale feature aggregation (34% receptive field increase). The proposed method achieves 98% accuracy (5.5% improvement over baseline YOLOv8) with 15 ms end-to-end processing time, validated through 5-fold cross-validation (± 0.36% std). The system addresses key agricultural challenges, including growth-stage variations and environmental noise, through texture consistency regularisation. A theoretical framework for temporal progression modelling is also proposed to guide future developments in time-series monitoring. An experimental evaluation on the Paddy Doctor dataset (10,407 images across 6 disease classes and a healthy class) demonstrates superior performance compared to existing methods. INT8 quantisation performed for mobile deployments on edge devices (model and compressed to 29 MB with an accuracy of 97.2% in comparison to Field testing showing 89% accuracy under challenging environmental conditions. This research work shows the advancement of agricultural AI through theoretical improvements and ready to use tools for farmers.